AI0-001 Implementing AI Solutions Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Use a rule-based system with keyword matching

  Why it's wrong here

  Keyword rules cannot generalise to the varied phrasing of 10,000 labelled tickets, so accuracy plateaus and maintenance grows with each new category. It is tempting for its transparency and zero training cost, and would be correct for a small, stable, strictly templated ticket set with unambiguous trigger words.

• ✗

  Use a prompt-based LLM with few-shot examples

  Why it's wrong here

  Few-shot prompting leaves the 10,000 labelled examples unused and yields inconsistent labels at scale, with no reproducible decision boundary. It is tempting because it needs no training pipeline, and would be correct where labelled data is scarce and categories are few, novel or rapidly changing.

• ✓

  Fine-tune a pre-trained text classification model

  Why this is correct

  Fine-tuning a pre-trained text classification model leverages the 10,000 labelled tickets to adapt existing language representations to the organisation's specific categories, satisfying the stem's supervised categorisation requirement. This outperforms training from scratch on a small dataset and avoids the cost and latency of prompt-based large-model inference.

• ✗

  Train a custom neural network from scratch

  Why it's wrong here

  Training a custom neural network from scratch requires tens of thousands of labelled examples per category to avoid overfitting, yet the dataset contains only 10,000 tickets total, which is insufficient for reliable generalisation. This approach is tempting because it offers full architectural control for highly specialised or novel classification tasks, and would be correct if the organisation possessed orders-of-magnitude more data and lacked a pre-trained language model.